Skip to content
Back to search results

Linear Quadratic Gaussian

Introduced by
Rudolf Kalman
Year
1960
Type
algorithm
Family
Stochastic Control
Font size

What does it do?

The Linear Quadratic Gaussian (LQG) controller combines the Linear Quadratic Regulator (LQR) with a Kalman Filter to handle stochastic systems with measurement noise and process noise.

Developed by Kalman and later formalized by Athans and others, LQG is the natural stochastic extension of LQR and remains the gold standard for optimal linear control under noise, with applications spanning spacecraft, aircraft autopilot, and industrial process control.

The intuition behind the method

LQG solves the problem of controlling a linear system that is corrupted by Gaussian noise, where only partial and noisy measurements of the state are available. The solution separates into two parts: a state estimator (Kalman Filter) that reconstructs the state from noisy measurements, and an optimal controller (LQR) that acts on the estimated state. This separation principle—the Kalman Filter and LQR are computed independently, then combined—is remarkable: the optimal LQG solution is simply applying the LQR gain to the Kalman-estimated state. This elegant decomposition enables modular design and analysis.

History

Kalman published the filtering algorithm in 1960 and extended it to the stochastic LQ problem shortly after. Athans provided a comprehensive treatment in 1971, establishing the separation principle and robustness properties. Kwakernaak and Sivan published a definitive textbook in 1972 covering both theory and applications. The introduction of the Extended Kalman Filter (EKF) by Bucy and others in the mid-1960s extended LQG to nonlinear systems. LQG remains central to modern control theory and has inspired related methods like H-infinity control and robust control.

When to use it

Use LQG when you have a linear system subject to Gaussian process and measurement noise, when full state measurement is unavailable or expensive, and when you need optimal control under uncertainty. LQG is ideal for applications like spacecraft attitude control, inertial navigation, and industrial systems with uncertain disturbances. Avoid LQG if the system is strongly nonlinear (use extended or unscented Kalman filter variants), if noise is non-Gaussian, or if measurements arrive at irregular intervals.

Limitations and considerations

  • Assumes Gaussian noise; real disturbances are often non-Gaussian with outliers or multi-modal distributions.
  • Requires accurate knowledge of process and measurement noise covariances (Q, R); misspecified noise statistics degrade performance.
  • Linear assumption; nonlinear systems require approximations (EKF, UKF) that lose optimality guarantees.
  • Assumes linear measurements; nonlinear measurement relationships require extended or unscented variants.
  • Cannot handle hard constraints; soft quadratic penalties only.

Source conditions do not replace independent scientific review or validation by an analysis engine.

Strengths

  • Provably optimal for linear systems with Gaussian noise; minimizes the expected quadratic cost.
  • Separation principle: Kalman Filter and LQR can be designed independently, simplifying analysis and debugging.
  • Provides both optimal state estimation and optimal feedback control in a unified framework.
  • Robustness properties inherited from both the Kalman Filter (noise rejection) and LQR (stability margins).
  • Real-time implementation feasible; both Kalman filter update and LQR gain application are O(n^3) and O(n^2) respectively.

Applications

  • Aerospace: Spacecraft attitude estimation and control, inertial navigation systems, autopilot design for aircraft.
  • Robotics: Sensor fusion in mobile robots, SLAM (Simultaneous Localization and Mapping) with measurement noise.
  • Automotive: Odometry fusion with GPS, engine control, adaptive suspension damping.
  • Power systems: Load frequency control, state estimation in smart grids with PMU (phasor measurement unit) data.
  • Chemical processes: Reactor temperature and composition control, distillation column optimization under measurement noise.

How it works

How it works

1

Run a Kalman Filter to recursively estimate the system state from noisy measurements using the system model and noise covariances.

2

Compute the optimal LQR feedback gain using the system dynamics and quadratic cost matrices.

3

At each sampling time, measure the noisy output and update the Kalman Filter estimate.

4

Apply the control input u = −K * x_hat (where x_hat is the estimated state) and repeat with new measurements.

Common pitfalls

  • Tuning noise covariances Q and R by trial-and-error instead of estimating them from data; this leads to suboptimal filtering or excessive noise amplification.
  • Ignoring the separation principle and trying to jointly optimize estimation and control, leading to unnecessary complexity.
  • Initializing the Kalman Filter state estimate incorrectly; poor initial conditions can slow convergence.
  • Assuming the system is observable and controllable without verifying; if not, the Kalman Filter or LQR may be unstable.

Frequently asked questions

What is the separation principle and why is it so important?

The separation principle states that for LQG, the optimal state estimator (Kalman Filter) and the optimal controller (LQR) can be designed independently, and combining them is optimal. This means you can first design a Kalman Filter ignoring control, then design an LQR controller ignoring noise, and apply both together without loss of optimality. This simplifies design and enables modular system development.

How do I estimate the process noise covariance Q and measurement noise covariance R?

The most principled approach is to estimate Q and R from data using maximum likelihood or covariance matching methods. Alternatively, start with rough estimates based on sensor specifications and disturbance magnitudes, then tune empirically by monitoring filter residuals (they should be white and uncorrelated). If residuals are correlated, increase Q or R; if the filter diverges, reduce them.

What happens if my system is nonlinear?

For mildly nonlinear systems, linearize around the nominal trajectory and use the Extended Kalman Filter (EKF) combined with LQR designed for the linearized system. For strongly nonlinear systems, consider the Unscented Kalman Filter (UKF) or particle filters, though these lose the optimality guarantee. Model Predictive Control with nonlinear models is another alternative.

Can I use LQG with time-varying systems?

Yes, the Kalman Filter and LQR both have time-varying variants. The time-varying Kalman Filter solves a Riccati equation backward in time from a terminal cost. The time-varying LQR solves a Riccati equation backward in time as well. Both are more computationally expensive than the steady-state versions but enable handling of trajectory tracking and slowly time-varying dynamics.

Related methods

Sources

  1. Kalman, R. E. (1960). A new approach to linear filtering and prediction problems. Journal of Basic Engineering, 82(1), 35-45.

    Open DOI reference
    Copied.
  2. Athans, M. (1971). The role and use of the stochastic linear-quadratic-gaussian problem in control system design. IEEE Transactions on Automatic Control, 16(6), 529-552.

    Open DOI reference
    Copied.
  3. Kwakernaak, H., & Sivan, R. (1972). Linear Optimal Control Systems. Wiley-Interscience.

    Open publication reference
    Copied.